---
title: "Adaptive Multilayer Perceptron"
output: 
  flexdashboard::flex_dashboard:
    storyboard: true
    social: menu
    source: embed
  
    
---


```{r setup, include=FALSE}
library(flexdashboard)
```

### Apresentation.

![alt](/Users/Fernanda/Downloads/17.png)

### Back to version 1




![alt](/Users/Fernanda/Downloads/1.png)







###  What is the objective again ?



![alt](/Users/Fernanda/Downloads/3.png)


### Version 2.



![alt](/Users/Fernanda/Downloads/adaptive_mlp.png)

### Change the number of neurons.



![alt](/Users/Fernanda/Downloads/4.png)

### Change the number of layers.


![alt](/Users/Fernanda/Downloads/change.png)



### Implementation challenges.


![alt](/Users/Fernanda/Downloads/5.png)

***

1. Add more layer at will
2. Generalization



### Output Layer.


![alt](/Users/Fernanda/Downloads/6.png)


### Hidden Layer 2.


![alt](/Users/Fernanda/Downloads/7.png)



### Hidden Layer 1.

![alt](/Users/Fernanda/Downloads/8.png)



### Perception.

![alt](/Users/Fernanda/Downloads/9.png)


***
1. Cross-entropy loss:
    - each epoch there’s a error loss

2. Batch accuracy:
    - for each 10 rows of data there will be a accuracy among them

3. Convergence rate:
    - for each 10 epochs the rate of convergence (error loss vs number of epochs) will be calculated



### Demo

![alt](/Users/Fernanda/Downloads/13.png)

### Next Steps

![alt](/Users/Fernanda/Downloads/14.png)


### Search of best configurations.


![alt](/Users/Fernanda/Downloads/10.png)

***

1. Hypothetical case

  - ANN with 4 layer maximum (1 inputs 2 hidden 1 output)
  - 3 variants of number of neurons for each hidden layer
  - 2 variants of number of layers (with 1 or 2 hidden layers)
  - 3 types of activations functions
  - 243 possibilities with 4 layers
  - 27 possibilities with 3 layers
  - Total of 270 possibilities



### Finding the best architecture.

![alt](/Users/Fernanda/Downloads/Which-One-is-Best-for-You.png)



***

1. Exhaustive search is not a good idea
2. Intuitively decide the number of neurons, layers and activations functions
or maybe 
3. We can use a meta heuristic to search in the search space 
	genetic algorithm ?



### Genetic Algorithm Search.

![alt](/Users/Fernanda/Downloads/15.png)

***

1. Each component is equivalent to a gene
2. Each configuration is a individual
3. The change of configuration is related to the crossing of two individuals or a mutation
4. Genetic operations with MEMORY!!



### Questions?


![alt](/Users/Fernanda/Downloads/12.png)